China AI Open-Weight Monetization: Winners Take More, Raise Zhipu AI Target, Cut MiniMax Target

Overview

This report discusses the open-weight monetization path for Chinese large language model vendors. JPMorgan says investors typically view open-weighting as monetization leakage because third parties can deploy models outside one-party APIs; but the more important question is whether vendors can turn broader distribution and weaker access control into paid conversion. The report argues this depends heavily on model capability: stronger models can scale adoption through cloud providers, inference platforms, and private deployment while capturing value through official support routes, licensing, marketplace SaaS, revenue sharing, deployment support, and enterprise co-selling. Weaker models are more likely to face price compression, substitution, and traffic rerouting.

Core views

The core view is that open-weight commercialization is forming a 'winners take more' framework. Open-weighting itself is simply releasing checkpoints, while official APIs are an ongoing optimized product that can differentiate on model freshness, prompt caching, long context, tool calling, throughput, latency, reliability, and enterprise SLA. Cloud providers and inference platforms can choose to self-deploy to retain more gross margin, or partner with model vendors to obtain better version synchronization, technical support, and endpoint quality. For high-value coding, agent, long-context, and enterprise production workloads, official APIs or official supported endpoints remain commercially relevant.

Analysis framework

The report follows a core narrative of business model framework, endpoint quality differences, cloud-channel distribution, workflow-product migration, and company valuation adjustments, comparing monetization capability of open-weight models across proprietary APIs, third-party deployment, cloud marketplaces, inference platforms, and proprietary workflow products. At company level, it places Zhipu AI and MiniMax in a common framework and assesses whether GLM-5.2 has stronger open-weight monetization optionality, and whether MiniMax M3 shows enough model differentiation and pricing power.

Methodology notes

• Monetization FrameworkOpen-Weight Winners Take More

  Model capability determines whether open-weighting is a distribution lever or monetization leakage

  Strong models can expand adoption with open-weighting while still capturing revenue via official APIs, cloud partnerships, enterprise deployment, and workflow products; weaker models are more easily replaced by third-party lower-price endpoints.

• Channel EconomicsCloud Providers and Inference Platforms: Margin vs Quality Trade-off

  Self-deployment preserves margin, while official partnership improves quality

  Cloud platforms can self-deploy open-weight models and preserve more economics, but must assume burdens of quantization, batching, caching, version updates, security, and reliability; partnership routes reduce take rate but improve model freshness, endpoint quality, and enterprise trust.

• Product StrategyOne-Vendor Workflow Harness

  Shift from interchangeable API backends to high-stickiness workflow entry points

  If a model vendor has coding agents, workflow products, or enterprise tools, open-weight can act as a funnel entry, while proprietary products convert some adoption into steadier vendor usage and stronger customer relationships.

• API EconomicsEffective Cost and Endpoint Quality Trump Face Value Prices

  Cache efficiency, speed, and stability drive real payment potential

  The report emphasizes that customers buy production-grade model services, not just weights or listed prices; prompt caching, long-context stability, tool reliability, latency consistency, and task completion rates determine real value.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Zhipu AI / 2513.HK

  Positive beneficiary

  Strengths

  GLM-5.2 is viewed by the report as having a global competitive positioning; open-weighting can expand adoption and preserve monetization through official routes, the GLM-Turbo family, cloud-marketplace SaaS, enterprise deployment, and workflow products.

  Weaknesses

  Current valuation may already reflect the US$1bn year-end ARR guidance to a significant extent, and the revenue path still depends on sustained model leadership and paid conversion.

  Comparison

  Compared with MiniMax, the report believes Zhipu is more likely to gain distribution and monetization optionality under the open-weight framework.

  Risks

  If GLM-5.2 does not represent a company-specific capability step-up, or if competitive models like Kimi K3 and DeepSeek V4.1 narrow the gap, the target price and overweight logic would come under pressure.

• MiniMax Group Inc - H / 0100.HK

  Neutral under pressure

  Strengths

  MiniMax M3's official route is said to have quality advantages in cache hit rate, effective input cost, usage concentration, and speed relative to some third-party routes.

  Weaknesses

  The report argues M3 has not yet shown enough model differentiation or model-driven pricing power, and open-weighting may accelerate price comparison and substitution.

  Comparison

  Compared with Zhipu, MiniMax has weaker evidence of paid conversion in this framework, so its target price was lowered and rating kept Neutral.

  Risks

  If third-party endpoints offer lower prices at similar quality, or if users regard M3 as substitutable, API economics and workflow conversion would be pressured.

• DeepSeek V4 Pro

  Reference case

  Strengths

  The official route forms a pricing moat through lower list pricing and caching economics, with significantly lower cost than some third-party routes in a simplified workload.

  Weaknesses

  As a reference case rather than a core rating coverage name, endpoint quality and commercial capture continue to depend on official ongoing optimization and channel strategy.

  Comparison

  The case shows that even open-weight strong models can retain commercial relevance through the effective cost and quality of official APIs.

  Risks

  If third-party deployments catch up in caching, speed, and quality, the official-route advantage may narrow.

• Cloud providers and inference platforms

  Distribution channels and competitive channels

  Strengths

  Open-weight expands model availability, and platforms can reach global customers through channels such as AWS, Azure/Foundry, Cloudflare, and OpenRouter.

  Weaknesses

  Self-deployment requires taking on model maintenance, quantization, batching, caching, updates, security, and reliability responsibilities.

  Comparison

  Official partnership routes are generally higher and more stable in quality but share economics, while self-deployment routes offer higher margin but greater quality risk.

  Risks

  If customers prioritize price and availability, traffic can move from official APIs to third-party endpoints; if third-party experiences are poor, traffic may return to official routes.

Key data

• Zhipu AI target priceHK$2,000Dec-26 target price was raised from HK$1,800 to HK$2,000, with rating maintained at Overweight.
• MiniMax target priceHK$300Dec-26 target price was reduced to HK$300, with rating maintained at Neutral.
• DeepSeek V4 Pro official-route cost caseabout US$24-41 vs third-party-route US$85-196Based on a simplified monthly workload of 100M effective input tokens and 20M output tokens, the official route benefits from lower list pricing and stronger caching economics.
• Adjusted FY26E EPS for Zhipu AIRmb (8.20)The table shows prior value Rmb (8.29), current Rmb (8.20), adjusted about -1.0%.
• Adjusted FY27E EPS for Zhipu AIRmb (6.71)The table shows prior value Rmb (7.45), current Rmb (6.71), adjusted about -10.0%.
• Zhipu AI revenue forecastFY25A Rmb 724mn; FY26E Rmb 4,715mn; FY27E Rmb 12,324mn; FY28E Rmb 33,485mnThe report's table shows high revenue growth, while profitability is still in a ramp-up phase.
• MiniMax revenue forecastFY25A US$79mn; FY26E US$383mn; FY27E US$830mn; FY28E US$1,652mnThe report's table shows rapid revenue growth, but adjusted EBIT, EBITDA, and net profit remain negative.
• Adjusted FY26E EPS for MiniMaxUS$ (1.36)The table discloses FY26E adjusted EPS as negative.
• Adjusted FY27E EPS for MiniMaxUS$ (2.92)The table discloses FY27E adjusted EPS as a larger loss.

Impact & implications

For investment implications, open-weighting should not be interpreted simply as damage to API revenue, but rather as a shift in revenue structure: model access revenue may be pressured, while broader distribution can create opportunities for paid inference, cloud channels, enterprise deployment, and workflow products. The market is likely to prefer vendors with sustained model leadership, official endpoint quality, cloud-platform prioritization, and proprietary workflow capability. Zhipu AI's investment logic depends on whether GLM-5.2 and subsequent GLM-5.5/6 can sustain capability leadership and expand monetization optionality; MiniMax's key challenge is whether M3 can prove differentiation, pricing power, and paid conversion.

Risks

• Open-weight lowers access control, allowing third-party cloud platforms, API aggregators, and technically capable enterprises to self-deploy, which can divert traffic that might otherwise flow to official APIs.
• Simple chat, summarization, translation, and basic Q&A tasks with low SLA are easier to price-compare and switch among providers, and mature intelligent-capability pricing may continue to compress.
• If model capability is insufficient or differentiation is not clear, open-weight can accelerate substitution rather than improve paid conversion.
• Cloud provider self-deployment may lead to version lag, unstable endpoint quality, weaker caching efficiency, or reliability shortfalls, affecting the user experience of the same-named model.
• Proprietary workflow products face high localization barriers in China, where Tencent WorkBuddy, ByteDance Trae, and Alibaba Qoder already have entry and distribution advantages.
• Zhipu AI's bull case relies on GLM-5.2, GLM-5.5/6, and similar models continuing to maintain capability leadership; if competing models catch up, valuation and pricing power could come under pressure.
• If MiniMax cannot demonstrate M3 differentiation, speed, and conversion advantages, its target price and revenue quality may remain under pressure.

What to watch

• Whether GLM-5.2 represents a company-specific capability step-up for Zhipu AI.
• The relative model performance of Kimi K3, DeepSeek V4.1, and GLM-5.2.
• Whether GLM-5.5/6 can sustain or expand the capability gap.
• The distribution quality and commercial terms of open-weight models on overseas platforms such as AWS, Azure/Foundry, Cloudflare, and OpenRouter.
• The gap between official APIs and third-party deployments in cache hit rate, effective input cost, speed, long-context stability, and tool-calling reliability.
• Whether proprietary workflow products like Zhipu Z Code, MiniMax Code, and Kimi Code can build user stickiness and paid conversion.
• Whether enterprise customers migrate from cloud-storefronts, inference layers, or aggregators into official APIs, quality endpoints, or workflow products.
• Whether MiniMax M3 shows clearer evidence of model leadership, pricing power, or adoption in high-value workloads.